一种改进CycleGAN的低照度图像增强网络

A Low-light Image Enhancement Network to Improve CycleGAN

为了提高CycleGAN对低照度图像增强后的细节分辨能力,提高图像整体的视觉质量,提出了一种改进CycleGAN的低照度图像增强算法。该网络的生成器由低光照增强模块和亮度均衡处理模块组成,用以学习低照度图像到正常照度图像的特征映射。以多尺度卷积和残差空洞卷积构建基于U-Net的低光照增强模块,提高网络对增强后图像细节信息的恢复能力;以全卷积构建亮度均衡处理模块,使图像亮度分布均匀并且提高图像视觉质量;以Patch-GAN作为网络的判别器,利用映射为一个N×N的矩阵来提高判别器对细节信息的分辨能力,并且增强网络收敛能力。验证结果表明,相较于其他对比方法,所提算法增强后的图像拥有更好的主观视觉效果和图像细节,同时在客观指标上的结果也优于其他方法。

In order to improve CycleGAN’s capability to distinguish the details of enhanced low-light images and improve the overall visual quality of image,a low-light image enhancement algorithm to improve CycleGAN is proposed.The generator of the network is composed of a low-light enhancement module and a brightness equalization processing module to learn the feature mapping from low-light images to normal-light images.Among them,multi-scale convolution and residual hollow convolution are used to build a U-Net-based low-light enhancement module in order to improve the network’s capability to restore the enhanced image detail information;and a brightness equalization processing module is built with full convolution to make the image brightness distribute uniformly and improve image visual quality.At the same time,Patch-GAN is used as the discriminator of the network,and an mapping N×N matrix is used to improve the discriminator’s capability to distinguish detailed information and enhance network convergence capability.Verification results show that as compared with other comparison methods,the image enhanced by the proposed algorithm has better subjective visual effects and image details,and the results on objective indicators are also better than other methods.